The promise of automated manufacturing for composite parts in the aircraft industry has not yet been fully realized. Though the safety-driven standards for high-quality structures are understandable, they are currently being met via manual visual inspection — i.e., an inspector must stop the machine and visually inspect the layup before proceeding.

ZAero is a European Commission (EC) funded project within the Horizon 2020 research framework which aims to increase composite parts production efficiency by 30-50% via an inline, automated inspection system. But its goes even further, encompassing part simulation and decision support tools for process chain operators.

Project leader: Profactor

The idea for this project came from Profactor (Steyr, Austria), a non-university R&D center with two main activities:

1. Micro and nano manufacturing
2. Industrial assistance systems such a human robot collaboration (cobots) and quality assurance using machine vision and machine learning.

“We are coordinator of the ZAero project,” says Profactor manager for machine vision, Dr. Christian Eitzinger. “Essentially, the project is our idea. We have worked with the aerospace industry for a while and seen problems with quality control in production of CFRP parts. From Airbus we learned there were still issues with inspection, especially with AFP, so we had the idea of a larger project that covers the whole production process.”

The whole production process, says Eitzinger, “means starting from layup and processes that place material into a mold, to curing processes, focusing on how to integrate quality control and the resulting rework decision and process chain.”

He explains that ZAero is just one in a large cluster of projects — see 4ZDM — within the field of zero defect manufacturing, not just for aerospace but in different fields (e.g., micro-manufacturing in aerospace as well as automotive).

Current State of the Art (SOTA)

ZAero Solution

SOURCE: Profactor
 

Sensors and machine learning

The sensors ZAero is using for layup have been developed in-house. It is also pursuing sensors developed and recently qualified by InFactory Solutions for MTorres AFP equipment. “Our experiments are currently based on dry fiber placement with MTorres and FIDAMC,” Eitzinger adds.

One challenge is that sensors need to be integrated into the machine. “But there is little space to do this in the AFP heads,” Eitzinger points out. “It requires delicate design.” Another issue is development of sensors for AFP onto curved surfaces. “In the end, it’s an economic issue,” he explains. “If you want to build a sensor to scan a surface, it has a certain depth of view. If the part curvature is high, then the area you can scan within your depth of view is smaller. If you want to scan a larger area, you need more scanners or the ability to scan in more angles (e.g., 2 images with an offset of 10°).” This would be analogous to the use of phased array transducers in large-area ultrasonic testing (UT) systems.

Though machine learning is a Profactor expertise, developing the upper level decision support technology for the ZAero solution presents certain issues. Eitzinger explains, “The standard approach when you use a machine learning system is to train the system to reach a certain level of performance before implementing it in production. To do this, we train the system offline and then “freeze” it —  putting it into production with no further changes. However, with ZAero, we are pursuing a newer approach, trying to make use of the data stream from the production line to continuously improve the system’s performance. This is online training. The challenge, then, is that the system’s machine learning is changing all the time and perhaps not always in the right direction. Thus, one of the issues is how to assess and manage this online machine learning in order to keep the process chain stable.”

Simulation

This part of the project is led by Dassault Systèmes. “Their current focus is more on the layup process,” says Eitzinger. “If there are defects, what do you do? Instead of using the current set of complex rules, their approach is to acquire data from how the part is actually manufactured and use this in FEA models to predict the part’s performance. You simulate the defect using real data and predict what impact it will have. Can it stay or does it need to be removed/addressed?”

Stay tuned for our next blog on ZAero, which will include an in-depth discussion about the resin infusion flow front sensors being developed and resin state monitoring during cure.